resume-screening

name	resume-screening
description	Structured, bias-aware screening of resumes against a job description, producing per-candidate scores and a ranked shortlist.
when_to_use	the user asks to evaluate, score, rank, or compare resumes/CVs for a specific role

Resume Screening

A repeatable workflow for turning a pile of resumes plus a job description into a defensible, scored shortlist.

Preconditions (do not start without both)

1. The resumes — file paths the user has supplied (pdf/docx/txt). Read each with your document-reading tool.
2. The job requirements — a concrete role spec. If the user only gives a title, ask for must-have skills, years of experience, and any hard filters (location, clearance, degree) before scoring.

Scoring procedure

Score every candidate on the five dimensions in rubric.md (load that file with your file tool now). For each dimension give a 1–5 score and one line of evidence quoted from the resume — never a bare number.

1. Extract the candidate's relevant facts per dimension.
2. Apply the rubric anchors strictly; when evidence is missing, score low and say "not evidenced", do not infer.
3. Compute the weighted total with the bundled scorer — don't do the arithmetic by hand. Run: run_script("<resource_dir>/score.py", ["{\"core_skill_match\": 4, \"relevant_experience\": 3, \"demonstrated_impact\": 5, \"trajectory\": 3, \"communication\": 4, \"hard_filter_pass\": true}"]) It returns the weighted total and recommendation band deterministically.
4. Flag any hard-filter failures explicitly (they cap the recommendation at "no", regardless of score).

Bias guardrails

• Ignore name, gender, age, school prestige, and photos. Score skills and demonstrated outcomes only.
• Treat employment gaps neutrally unless the role specifically requires continuity.

Output

Produce a table (one row per candidate): name/id, the five dimension scores, weighted total, hard-filter pass/fail, and a one-sentence recommendation (strong yes / yes / maybe / no). If the user wants a file, write the table to the project directory as CSV and return its path.